An Unsmooth Feature Variation and Dynamic Graph Structure Update GCN for Machinery Fault Diagnosis

Fault detection of high-speed high-power diesel engines is highly challenging. Currently, multi-channel data is applied to fault detection of high-speed and high-power diesel engines. A common multi-channel data fault diagnosis method is the graph neural network (GNN), which can represent the connection relationship of different sensors as a graph and then diagnose the mechanical equipment. However, ordinary GNN suffers from the phenomenon that node features tend to be over-smoothed, which reduces the saliency of node features, as well as the need to set the linking method of nodes artificially, and the model cannot automatically correct the insufficiently reasonable graph structure, which leads to insufficiently stable diagnostic results. Therefore, we propose an unsmooth feature variation and dynamic graph structure update graph convolutional network (UDGCN). The network combines the improved feed forward network (FFN), a node feature variation layer for increasing the diversity of node information, and an improved graph convolutional network (GCN) which combined a dynamic graph structure update layer after a normal GCN for adaptively adjusting the way nodes are connected in the k nearest neighbors (KNN) graph. In the end, the datasets of misfire faults of diesel engines are tested and analyzed by training models, which show that the proposed method has higher diagnostic accuracy and stronger adaptability to the working conditions. Otherwise, this paper analyzes the interpretability of the model by the GraphLIME method with improved sampling method, which achieves certain interpretation effect and enhances the credibility of the model.